feat(models): port FastVLM (FastViTHD vision + Qwen2 text)

[inureyes]

Summary

Add FastVLM (Apple, llava_qwen2 / fastvlm): a FastViTHD hybrid convolutional/transformer vision encoder feeding a stock Qwen2 text decoder through an mlp2x_gelu MLP projector.

• Vision tower (src/vision/encoders/fastvlm.rs): the FastViTHD tower runs entirely on channels-last maps. A 3-block conv stem, three RepMixer stages (depthwise token mixing plus a BatchNorm ConvFFN with a layer scale), two attention stages (channel LayerNorm, head_dim 32 self-attention over the flattened map, dual layer scales), large-kernel PatchEmbed downsamples, RepCPE position encoders, and a squeeze-excite conv_exp head. The final (16, 16, 3072) map is flattened to 256 tokens. Reuses Conv2dLayer (grouped conv) and inference BatchNorm from existing modules.
• Projector: reuses MLPProjector (mm_projector.0 / .2 renamed to linear_1 / linear_2); rejects any mm_projector_type other than mlp2x_gelu.
• Processor (src/vision/processors/fastvlm.rs): pad-to-square, bicubic resize to 1024, rescale, per-channel normalize, channels-first (B, 3, 1024, 1024).
• Prompt / runtime: the <image> placeholder is the fixed -200 sentinel (not a vocabulary token). A dedicated runtime arm re-tokenizes the rendered prompt, splices one sentinel per image, expands each to 256 tokens, preprocesses images, and merges through the generic VisionModule (LLaVA scatter). The -200 id survives the embedding lookup (mx::take) and is overwritten by the scatter.
• Loader (src/loading/vlm_fastvlm.rs): text reuses Qwen2 unchanged (qkv biases auto-load from weight presence). Handles both the genuine (apple/FastVLM-0.5B, channels-first convs, (C,1,1) layer scales, patch_embed.<n>) and converted (mlx-community/FastVLM-0.5B-bf16, already channels-last) layouts; the conv layout is decided once from an unambiguous depthwise probe. Drops the classification head, BatchNorm counters, and the tied lm_head.

Validation

Real checkpoint mlx-community/FastVLM-0.5B-bf16:

• Text-only: "What is the capital of France?" produced "Paris."
• Shapes image (256 image tokens): produced "three distinct geometric shapes: a red circle, a blue square, and a green triangle. The red circle is positioned in the upper left corner ... the blue square is located in the upper right corner. The green triangle is situated in the lower center." Colors, shapes, and positions are all correct.

6 co-located unit tests and 4 real-model parity tests (tests/fastvlm_parity.rs) pass. Release build (metal,accelerate), cargo fmt, and cargo clippy --all-targets -- -D warnings are clean.

Out of scope

Quantized FastVLM conversions, LoRA on the vision tower, non-mlp2x_gelu projectors, the classification head, video, and dynamic resolutions (the model is fixed 1024x1024 pad-to-square).

Drive-by

Ignore in (cxx unsound advisory). mlxcel does not use that macro, so the path is unreachable; the advisory was published after the last merge and fails cargo-deny on every PR until ignored.

Closes #538